Reyes：一个从0到1开始训练的多模态大模型（技术报告） 原创

class ReyesModel(PreTrainedModel):
    config_class = ReyesConfig
    main_input_name = 'pixel_values'
    _supports_flash_attn_2 = True
    _no_split_modules = ['InternVisionModel', 'Qwen2DecoderLayer']

    def __init__(self, config: ReyesConfig, vision_model=None, language_model=None, use_flash_attn=True):
        super().__init__(config)

        assert version_cmp(transformers.__version__, '4.44.2', 'ge')
        image_size = config.force_image_size or config.vision_config.image_size
        patch_size = config.vision_config.patch_size
        self.patch_size = patch_size
        self.select_layer = config.select_layer
        self.llm_arch_name = config.llm_config.architectures[0]
        self.template = config.template
        self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
        self.downsample_ratio = config.downsample_ratio
        self.ps_version = config.ps_version
        use_flash_attn = use_flash_attn if has_flash_attn elseFalse
        config.vision_config.use_flash_attn = Trueif use_flash_attn elseFalse
        config.llm_config._attn_implementation = 'flash_attention_2'if use_flash_attn else'eager'

        logger.info(f'num_image_token: {self.num_image_token}')
        logger.info(f'ps_version: {self.ps_version}')
        if vision_model isnotNone:
            self.vision_model = vision_model
        else:
            self.vision_model = InternVisionModel(config.vision_config)
        if language_model isnotNone:
            self.language_model = language_model
        else:
            if config.llm_config.architectures[0] == 'Qwen2ForCausalLM':
                self.language_model = Qwen2ForCausalLM(config.llm_config)
                # self.language_model = AutoLigerKernelForCausalLM(config.llm_config)
            else:
                raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')

        vit_hidden_size = config.vision_config.hidden_size
        llm_intermediate_size = config.llm_config.intermediate_size
        llm_hidden_size = config.llm_config.hidden_size

        self.mlp1 = nn.Sequential(
            nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
            nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_intermediate_size, bias=False),
            nn.GELU(),
            nn.Linear(llm_intermediate_size, llm_hidden_size, bias=False)
        )

        self.img_context_token_id = None
        self.conv_template = get_conv_template(self.template)
        self.system_message = self.conv_template.system_message

        if config.use_backbone_lora:
            self.wrap_backbone_lora(r=config.use_backbone_lora, lora_alpha=2 * config.use_backbone_lora)

        if config.use_llm_lora:
            self.wrap_llm_lora(r=config.use_llm_lora, lora_alpha=2 * config.use_llm_lora)

    def wrap_backbone_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
        lora_config = LoraConfig(
            r=r,
            target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'],
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
        )
        self.vision_model = get_peft_model(self.vision_model, lora_config)
        self.vision_model.print_trainable_parameters()

    def wrap_llm_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
        # Determine the target modules based on the architecture of the language model
        if self.llm_arch_name in ['Qwen2ForCausalLM', 'LlamaForCausalLM']:
            target_modules = ['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj',
                              'mlp.gate_proj', 'mlp.down_proj', 'mlp.up_proj']
        else:
            raiseNotImplemented
        lora_config = LoraConfig(
            r=r,
            target_modules=target_modules,
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            task_type='CAUSAL_LM'
        )
        self.language_model = get_peft_model(self.language_model, lora_config)
        self.language_model.enable_input_require_grads()
        self.language_model.print_trainable_parameters()

    def forward(
            self,
            pixel_values: torch.FloatTensor,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            image_flags: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict isnotNoneelse self.config.use_return_dict

        # image_flags = image_flags.squeeze(-1)
        input_embeds = self.language_model.get_input_embeddings()(input_ids)

        vit_embeds = self.extract_feature(pixel_values)
        # vit_embeds = vit_embeds[image_flags == 1]
        vit_batch_size = pixel_values.shape[0]

        B, N, C = input_embeds.shape
        input_embeds = input_embeds.reshape(B * N, C)

        # if torch.distributed.get_rank() == 0:
        #     print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}')

        input_ids = input_ids.reshape(B * N)
        selected = (input_ids == self.img_context_token_id)
        try:
            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
        except Exception as e:
            vit_embeds = vit_embeds.reshape(-1, C)
            print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
                  f'vit_embeds.shape={vit_embeds.shape}')
            n_token = selected.sum()
            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]

        input_embeds = input_embeds.reshape(B, N, C)

        outputs = self.language_model(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentinotallow=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        logits = outputs.logits

        loss = None
        if labels isnotNone:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
            shift_labels = shift_labels.view(-1)

            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        ifnot return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss isnotNoneelse output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentinotallow=outputs.attentions,
        )

├── coco
│   └── train2017
├── gqa
│   └── images
├── ocr_vqa
│   └── images
├── textvqa
│   └── train_images
└── vg
    ├── VG_100K
    └── VG_100K_2

<Image><image></Image> Give a brief description of the image. A group of airplanes flying through the sky.<|im_end|>

<|im_start|>system
Follow the user's instruction and answer questions.<|im_end|>
<|im_start|>user
<Image><image></lmage>
What activity are the people engaged in on the green field?<|im_end|>
<|im_start|>assistant
The people are engaged in playing a game of frisbee on the lush green field.<|im_end|>
<|im_start|>user
Is there a person wearing a fairy outfit in the image? If so, what are they doing?<|im_end|>
<|im_start|>assistant
Yes, there is a person wearing a fairy outfit in the image. They are walking in the field, likely
participating in the fun activities along with the other people.<|im_end|>
<|im_start|>user
What can you say about the spectator's location in relation to the houses?<|im_end|>
<|im_start|>assistant
The spectators are watching a match being played in a location close to the houses, possibly indicating that the field is within a residential area or park.<|im_end|>

import torch
from PIL import Image
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)


def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB'else img),
        T.Resize((input_size, input_size), interpolatinotallow=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=True):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image_file, input_size=448, max_num=6):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

import torch
from modelscope import AutoTokenizer, AutoModel
from PIL import Image
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)


def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB'else img),
        T.Resize((input_size, input_size), interpolatinotallow=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values


def preprocess_image(file_path, dynamic=True, max_num=6, image_size=448):
    try:
        if dynamic:
            return load_image(file_path, max_num=max_num).to(torch.bfloat16).cuda()
        else:
            img = Image.open(file_path).convert('RGB')
            transform = build_transform(image_size)
            pixel_values = transform(img)
            return torch.stack([pixel_values]).to(torch.bfloat16).cuda()
    except Exception as e:
        raise RuntimeError(f"Error processing image: {e}")


path = "Reyes-8B"

model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).eval().cuda()

# print(model)

tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
generation_config = dict(max_new_tokens=2048, do_sample=False)

# single-image single-round conversation
file_path = 'tmp.png'
pixel_values = preprocess_image(file_path, dynamic=True)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')

# pure-text conversation
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

问题： Who painted <image 1>?
选项： {'A':'Claude Monet', 'B':'Henri Matisse', 'C':'Andy Warhol','D': "Georgia O'Keefe"]
预测的答案： C
正确的答案： C

问题： Each situation below relates to an independent company's Owners' Equity. <image 1> Calculate the missing values of company 2.
选项： {'A': '$1,620', 'B': '$12,000', 'C': '$51,180', 'D': '$0'}
预测的答案： D
正确的答案： D

问题： A survey line ABC crossing a river at right angles cut its banks at B and C, as shown in Fig. 2.39. To determine the width BC of the river, the following operation was carried out.A 60 m long line BE was set out roughly parallel to the river. Line CE was extended to D and mid-point F of DB was established. Then EF was extended to G such that FG = EF. Line DG was extended to cut the survey line ABC at H. GH and HB were measured and found to be 40 m and 80 m, respectively.Find the width of the river.<image 1>
选项： {'A': '120 m', 'B': '122 m', 'C': '123 m', 'D': '121 m'}
预测的答案： A
正确的答案： A